Factor of factors: letting ML learn the blend

The first walkthrough built a factor by blending raw signals with weights the AI proposed. This one goes a level up: the inputs are existing factors, and instead of fixing the weights, a machine learning model learns them. The output is a factor of factors — a composite whose blend adapts as the model retrains.

1. The idea: stop guessing the weights

Hand-set weights are a snapshot of today's opinion. When momentum leads in one regime and value leads in the next, a fixed blend is wrong half the time. So the analyst asks for something that re-learns the mix:

"Combine our core style factors into one score, and let a model decide how much each matters — retrained as the data moves."

2. Pick the input factors

The inputs here are already-built factors, each with its own track record in the library — not raw signals:

3. The model learns the blend

The analyst picks a model class; the Signal Designer handles the rest. The options are deliberately boring and robust:

Ridge and elastic-net give you a transparent linear blend; lightGBM captures interactions between factors when a linear mix leaves signal on the table. The model is trained walk-forward: for each date D it sees only data available strictly before D, then publishes a score for D.

4. What the model learned

Because the blend is learned rather than assigned, the "recipe" is the model's own coefficients (for the linear models) or feature importances (for lightGBM) — stored with the factor so anyone can see why it weights what it does:

Average learned importance over the calibration window. These re-fit on every retrain — the blend is not frozen.

5. Validation, standalone then in the book

The learned factor is evaluated on its own, then as an FMP alongside the existing library — the same bar every factor clears:

The correlation-to-inputs number matters most here: a good factor of factors should be more than the average of its parts. If it just tracks its inputs, the model hasn't earned its complexity.

The audit trail comes standard

Every ML factor lands in the library with its full provenance: the input factors, the model spec and hyperparameters, the coefficients or importances per retrain, the calibration windows, and the lookahead audit. A raw idea becomes a tracked, reviewable factor — with an audit trail behind it — before a dollar is allocated. See what's available to build from in the factor catalog.